Hyperspectral Remote Sensing: What 200+ Bands Reveal That Multispectral Can't

The moment hyperspectral imaging clicked for me was during a geological mapping project in Oman. Multispectral Sentinel-2 data could distinguish "bright rock" from "dark rock" — carbonates from basalts, broadly. But when I processed PRISMA hyperspectral data from the same area, I could identify specific carbonate minerals: calcite versus dolomite versus magnesite. Each had a diagnostic absorption feature near 2.3 μm, but at slightly different wavelengths — differences invisible to Sentinel-2's broad SWIR bands but clearly resolved by PRISMA's 10 nm channels.

That level of specificity is what makes hyperspectral remote sensing fundamentally different from multispectral. It's the difference between seeing colors and identifying chemical compositions.

Multispectral vs. Hyperspectral: The Core Difference

Multispectral sensors (Sentinel-2, Landsat) capture 10-13 bands, each spanning 60-180 nm of the electromagnetic spectrum. They sample the spectrum at discrete, widely spaced intervals. This is like hearing a symphony through a wall — you get the general melody but lose the fine detail.

Hyperspectral sensors capture 200+ contiguous bands, each only 5-10 nm wide, producing a nearly continuous spectrum for every pixel. This is like sitting in the concert hall — every instrument is distinguishable.

Feature	Sentinel-2 (Multispectral)	PRISMA (Hyperspectral)
Bands	13	240
Band width	15-180 nm	~10 nm
Spectral range	443-2190 nm	400-2500 nm
Spatial resolution	10-60m	30m
Swath width	290 km	30 km
Data per scene	~800 MB	~5 GB

The narrow swath and large data volume are the practical trade-offs. Hyperspectral sensors sacrifice spatial coverage for spectral detail.

Current Hyperspectral Satellite Missions

PRISMA (Italy, ASI, 2019-present)

The PRecursore IperSpettrale della Missione Applicativa has been a game-changer for the research community:

• 240 bands from 400-2500 nm at ~10 nm spectral resolution
• 30m spatial resolution
• 30 km swath width
• Data freely available for registered users

EnMAP (Germany, DLR, 2022-present)

The Environmental Mapping and Analysis Program:

• 228 bands from 430-2450 nm
• 30m spatial resolution
• 30 km swath width
• Improved signal-to-noise ratio compared to earlier missions

EMIT (NASA, ISS, 2022-present)

Originally designed to map mineral dust sources, EMIT has become a powerful methane and CO₂ point-source detector:

• 285 bands from 380-2500 nm
• 60m spatial resolution
• 75 km swath width
• Mounted on the International Space Station

Upcoming: CHIME (ESA/Copernicus)

The Copernicus Hyperspectral Imaging Mission for the Environment, planned for 2028-2029, will be the first operational hyperspectral satellite:

• ~210 bands from 400-2500 nm at 10 nm resolution
• 30m spatial resolution
• ~130 km swath width — dramatically wider than current missions
• Systematic global coverage with ~11 day revisit

CHIME will transform hyperspectral from a research tool to an operational monitoring capability.

What Hyperspectral Can Do That Multispectral Cannot

Mineral Identification

This is the classic hyperspectral application. Minerals have diagnostic absorption features at specific wavelengths:

• Iron oxides: Absorption near 900 nm (goethite vs. hematite distinguished by exact position)
• Carbonates: Absorption near 2340 nm (calcite) vs. 2320 nm (dolomite)
• Clay minerals: Absorption doublet near 2200 nm (kaolinite vs. montmorillonite vs. illite)
• Sulfates: Absorption near 1750 nm and 2200 nm

Sentinel-2 can detect the presence of iron oxides (its SWIR band spans the absorption region) but cannot distinguish between goethite and hematite. Hyperspectral sensors can, because they resolve the exact absorption wavelength.

Crop Species and Stress Detection

Different crop species have subtly different spectral signatures, especially in the red-edge region (680-750 nm). Multispectral sensors have 1-3 bands here; hyperspectral sensors have 10-15.

More importantly, specific stresses produce diagnostic spectral changes:

• Nitrogen deficiency: Shifts the red-edge position toward shorter wavelengths
• Water stress: Deepens the 1450 nm and 1940 nm water absorption features
• Disease-specific signatures: Some fungal infections produce unique spectral patterns before visible symptoms appear

Water Quality Parameters

Hyperspectral data enables retrieval of specific water constituents:

• Chlorophyll-a concentration from the shape of the reflectance peak near 700 nm
• Cyanobacteria detection using the phycocyanin absorption feature near 620 nm
• Dissolved organic matter from the spectral slope in visible wavelengths
• Suspended sediment type (organic vs. mineral) from the spectral shape in red-NIR

Spectral Unmixing

In a 30m pixel, you might have a mixture of vegetation, soil, and shadow. Spectral unmixing decomposes this mixed spectrum into its constituent endmembers and their fractional abundances.

With multispectral data, unmixing is heavily under-determined (more endmembers than bands). With 200+ hyperspectral bands, the problem becomes over-determined, enabling reliable unmixing of complex scenes into 5-10 endmember materials.

Processing Challenges

Atmospheric Correction Is Critical

For multispectral analysis, approximate atmospheric correction is often sufficient — NDVI is relatively robust to atmospheric effects because it's a ratio. For hyperspectral analysis, atmospheric correction must be precise because you're comparing the exact shape of absorption features, and atmospheric water vapor and aerosols produce absorption features in the same spectral regions as your target materials.

Tools: ATCOR, FLAASH, and the newer ISOFIT provide hyperspectral-specific atmospheric correction.

The Hughes Phenomenon (Curse of Dimensionality)

With 200+ bands, many machine learning classifiers perform worse, not better, than with a well-chosen subset of 10-20 bands. This is because:

• Adjacent bands are highly correlated (adding redundant features)
• Training sample requirements increase exponentially with dimensionality
• Noise in individual bands accumulates

Solutions: Dimensionality reduction (PCA, MNF transform), band selection, or algorithms designed for high-dimensional data (SVM with RBF kernel, random forest).

Storage and Processing

A single PRISMA scene is ~5 GB. A regional study might involve dozens of scenes. Processing 200+ bands through atmospheric correction, geometric correction, and classification requires significantly more compute than equivalent multispectral workflows.

Getting Started with Hyperspectral Data

1. Register for PRISMA data access at ASI's portal (prisma.asi.it) — free for research
2. Register for EnMAP data at DLR's portal — also free
3. Start with EnMAP-Box — DLR's free QGIS plugin specifically designed for hyperspectral analysis
4. Use spectral libraries — USGS Spectral Library and ECOSTRESS Spectral Library provide reference spectra for minerals, vegetation, and man-made materials
5. Begin with a focused question — "Can I distinguish mineral X from mineral Y?" is answerable; "Map everything" is not

The field is at an inflection point. For decades, hyperspectral remote sensing was limited to expensive airborne campaigns and a handful of research satellites with narrow coverage. CHIME will change that, providing systematic global hyperspectral coverage for the first time. The analysis techniques developed on PRISMA and EnMAP data today will scale to global applications within a few years.